University of Toronto: APS360 Final Report – Group 21

TWEET SENTIMENT CLASSIFICATION USING CONVO-

LUTIONAL NEURAL NETWORKS

James Li

Student# 1007974248

jamesw.li@mail.utoronto.ca

Pierre Ishak

Student# 1007897457

pierre.ishak@mail.utoronto.ca

Tanmay Patel

Student# 1008279168

tanmayn.patel@mail.utoronto.ca

Nicholas Lupu-Muslea

Student# 1007896154

n.lupumuslea@mail.utoronto.ca

Link to GitHub: https://github.com/piroo8/Aps360

1 PROBLEM STATEMENT

In this day and age, where most everyday interactions are digitized, it is absolutely vital to be able

to extract essential information from said interactions and use them to your advantage. According

to an investigation conducted by McKinsey and Co.,(1) in 2020 following the COVID-19 crisis, it

was noted that the global digitization of customer interactions drastically accelerated from 36% in

December 2019 to 58% in July 2020 and is still growing to this day.

Given the context, the project goal is to create a tweet sentiment analysis classiﬁcation model using

a convolutional neural network (CNN) that would classify the tweet as one of three sentiments:

positive, negative or neutral. This model can then be used by other companies by transfer learning,

to analyse and classify text data to deduce quality insights regarding their products. Since social

media presence is so abundant, this tool will be great for marketing teams looking to gain insights

on product quality.

This is demonstrated by training the model using a large database with a variety of tweets and testing

it using a separate database, a variety of tweets mentioning Dell. It is important to discuss why deep

learning is necessary for such a task; the ability to learn hierarchical representations and model

complex relationships is what it a great candidate for the task. In addition, a CNN is easily able

to detect low-level features that include words or word combinations which are then used to learn

higher level features such as phrases and expressions with those words. What a CNN offers that is

different from an LSTM or RNN is that it does not require time-series data, appropriate for tweets

which do not necessarily follow ordinary grammatical and sequential structure. By leveraging these

deep learning techniques, the aim is to build a robust model that can accurately classify sentiments

in texts, enabling faster and objective-based analysis. Click the following for the illustration: 4

2 BACKGROUND AND RELATED WORK

2.1 SENTIMENT ANALYSIS USING DEEP LEARNING: FINANCIAL MARKET PREDICTION (1)

The report from McKinsey highlights that COVID-19 accelerated technological adoption, driving

companies across industries towards a permanent digital transformation. It emphasizes that organi-

zations which effectively leverage technology have gained a competitive edge, reshaping business

operations and customer interactions. Sentiment analysis using deep learning emerges as a pivotal

tool, enabling ﬁrms to understand and respond to evolving customer preferences, thus shaping their

strategies in this tech-driven landscape.

2.2 SENTIMENT ANALYSIS USING RECURSIVE NEURAL NETWORKS (RECNN) (3)

The report explores the signiﬁcant advantages of utilizing Deep Learning for Sentiment Analysis. It

delves into the effectiveness of deep neural networks in capturing intricate emotional nuances from

textual data. The study emphasizes the model’s enhanced ability to extract context-based sentiment,

contributing to more accurate sentiment classiﬁcation in various applications.

2.3 OPINION MINING (4)

The report ”Sentiment Analysis and Opinion Mining: A Survey” provides an in-depth examination

of sentiment analysis techniques, highlighting the growing prominence of deep learning in this ﬁeld.

It explores various methodologies for sentiment classiﬁcation, emphasizing the advantages of lever-

aging deep learning models such as neural networks. The survey underscores how deep learning

enables more accurate sentiment analysis by capturing complex linguistic nuances and contextual

information, leading to improved sentiment classiﬁcation performance.

University of Toronto: APS360 Final Report – Group 21

2.4 SENTIMENT ANALYSIS: TWITTER & COVID-19 (5)

The report explores the signiﬁcant advantages of employing deep learning techniques for sentiment

analysis. By utilizing advanced neural network architectures, the study demonstrates enhanced ac-

curacy in sentiment classiﬁcation tasks, particularly in capturing complex emotional nuances. The

ﬁndings highlight the potential of deep learning to optimize sentiment analysis processes across

various domains, underscoring its capacity to uncover deeper insights from textual data.

3 DATA PROCESSING

The ﬁrst data set that we used was a twitter data set from Kaggle. The reasoning for choosing

this data set was the sheer magnitude of the data containing over 60,000 tweets. In addition to its

magnitude, it also incorporated a breadth of tweets i.e. the tweets were from various sources such

as product reviews, movie reviews, gaming chats, etc.Link to first data set (train)

For the second data set, we also chose a dataset from Kaggle instead of collecting our own data since

the task at hand was complex and a large dataset was needed. Although, this time instead of choosing

a generic data set we decided to use a company directed data set to mimic real world scenario. We

considered ourselves to be a third party helping a company ﬁlter out tweets into different categories.

Hence, our choice of the Dell oriented data set as the test data. Link to second data set

(test)

3.1 DATA SOURCING, CLEANING & FORMATTING:

The data for this project was sourced from Kaggle (see above). While the dataset had received

a high quality and usability rating by other users on the website, our inspection revealed certain

aspects that required cleaning and reorganization to be fed into our model for optimal performance.

The following section discusses the different steps we took to prepare the data in a manner we

deemed suitable and easier for the model to use.The collected data underwent a series of cleaning

and formatting steps to ensure its suitability for training the model. The same formatting was applied

to the test data before being fed into the model to make predictions. The following is a summary of

the preprocessing steps applied to the data in order:

Type of Processing Description

Removing NaN,

NULL value rows

Rows with no data in the ”sentiment text” column were deleted from the

dataset as they were not useful for training.

Removing columns Remove all columns that added no sentiment value for analysis.

Normalization to

lowercase

All ”sentiment text” in the data was converted to lowercase to ensure

consistency and eliminate case sensitivity.

Removing punctua-

tion

Removed from the text to improve efﬁciency during tokenization, as well

as make tokenization easier.

Removing men-

tions (@) & URLs

Removed as they did not offer valuable information for sentiment classi-

ﬁcation.

Removing hashtags The ”#” symbol of hashtags was removed, retaining the meaningful tag.

Removing emojis All emojis were removed using both the emoji library and regular ex-

pressions. Emojis can be used to determine sentiments; however it is an

extremely difﬁcult challenge and not all tweets contained emojis.

Removing STOP

words

Non-informative words were removed to optimize sequence storage.

Some important words that we thought conveyed sentimental informa-

tion such as not and weren’t with negative connotation were kept.

Tokenization Converted tweets to tokens to feed the NLP model. Helps with feature

extraction, text segmentation, and vocabulary building.

Lemmatization

(over stemming)

Converts all words to base form depending on context. Normalization of

words improves text analysis and reducing features.

Standardizing la-

bels/classes

Removed rows with irrelevant label as did not prove beneﬁtial for train-

ing and second data set did not include this class

Labels to integers Conversion of labels to integer representations. Mapped negative → 0,

neutral → 1, positive → 2.

3.2 DATA STATISTICS

The statistics (entries) comparing pre- and post-processing: 74682 → 59526 training entries, 1000

→ 827 validation entries, 4 → 3 sentiments (see 1a, 1b).

3.3 PREPARING PROCESSED DATA FOR MODEL INPUT

Following data cleaning, the data was prepared for model input using functions from

torchtext.utils.data and torch.nn.utils.rnn. The tokened tweets were used to

University of Toronto: APS360 Final Report – Group 21

(a) Number of Entries in Each Class Pre-Processing. (b) Number of Entries in Each Class Post-Processing.

build a vocabulary based on frequency of words which was then used to convert the strings to a list

of integers indices. This proves to be vital for memory efﬁciency, maintaining ﬁxed-length-inputs,

(mini) batch processing, GPU acceleration, ease of transformation into tensors, and compatibility

of with the model. The tokened sequences of indices were then padded using pad sequence

to the maximum length of a tweet to ensure uniform length for efﬁcient training. The length of

each individual sequence before padding was also calculated and stored in a variable to avoid the

caveat of ’last output unit’ during batch processing as discussed in tutorial. Finally the labels were

converted to a tensor. Now, the three requirements for our input data were matched and ready for

data loading. TensorDatasets functionlaity was used to prepare a data set containing the tensor

versions of the padded sequences, the length of each sequence before padding, and the truth labels.

DataLoader was then used to load the data set into a loader while batching with drop last

being true. The last few entries were dropped to ensure that every batch was uniform. The data is

now ready to be used by the model.

4 MODEL ARCHITECTURE: TWEET SENTIMENT ANALYSIS USING CNNS

4.1 MODEL ARCHITECTURE

Our primary goal was to develop a CNN model capable of accurately classifying tweet sentiments

into predeﬁned categories. After experimenting with various architectures, including LSTM and

GRU, and a bidirectional LSTM and a GRU, we determined that the CNN architecture provided

the best performance. Notably, our selected CNN model achieved outstanding results, boasting a

training/validation accuracy of 97.6%. Impressively, it also attained a 50% accuracy on previously

unseen data from a distinct Dell tweets dataset.

4.2 CNN MODEL CLASS

To implement the sentiment analysis model, we introduced a custom CNN Model class, inheriting

from the nn.Module framework. The architecture consists of the following key components:

Component Description

Embedding

Layer

The model starts with an embedding layer that maps input tokens (words) to dense

vector representations of embedding dim dimensions.

Convolutional

Layers

Two consecutive convolutional layers, conv1 and conv2, are applied. conv1

employs a kernel size of 3, producing 128 output channels. Subsequently, conv2

reduces the channels to 64 using the same kernel size. ReLU activation functions

follow each convolutional layer.

Max-

Pooling

Layers

We employ max pooling with a kernel size of 3 to downsample feature maps,

thereby reducing their spatial dimensions.

Fully Con-

nected

Layers

Our architecture includes two fully connected layers, fc1 and fc2. Following

convolution and pooling, fc1 reduces ﬂattened feature maps to 128 dimensions,

while fc2 generates the ﬁnal output logits with num classes dimensions.

Dropout To mitigate overﬁtting, dropout is applied after fc1 with a dropout rate of

dropout percent.

University of Toronto: APS360 Final Report – Group 21

4.3 FORWARD PASS

In the forward method, input data undergoes the following steps:

Step Description

Embedding Input tokens traverse the embedding layer, yielding dense vector representations.

Permutation We permute the dimensions of the embedding output to prepare for convolutional

operations.

Convolution

and Pooling

The data undergoes convolution using the conv1 layer, followed by max pooling.

This process is repeated with conv2 and an additional max pooling layer.

Flattening Pooled feature maps are ﬂattened into a vector.

Fully Con-

nected Layers

The ﬂattened features progress through fc1 and fc2, culminating in the ﬁnal out-

put logits.

Dropout and

Output

The output of fc1 is subjected to dropout, ultimately yielding sentiment logits via

fc2.

4.4 HYPERPARAMETER OPTIMIZATION

Our pursuit of optimal model performance involved meticulous hyperparameter tuning using the

hyperopt library. We deﬁned a parameter search space encompassing learning rates, dropout

rates, optimizer choices, and epochs, enabling efﬁcient exploration of diverse conﬁgurations.

Objective Function: The objective function was designed to optimize the negative accuracy. We

recorded the top 10 conﬁgurations based on accuracy for further analysis.

Top Hyperparameters: After rigorous optimization, we identiﬁed the following optimal hy-

perparameters: {learning rate: 0.001; dropout: 0.2; optimizer: Adam;

epochs: 10}

5 BASELINE MODEL

PyCaret, a machine learning library, was used to implement a baseline model for sentiment analysis

on the tweet dataset, simplifying the process by automating various tasks, including data preprocess-

ing, feature engineering, hyperparameter tuning, and model comparison. With PyCaret, we tested

multiple classiﬁcation algorithms on the tweet dataset to identify the best-performing model by ac-

curacy. The LightGBM classiﬁer emerged as the top performer based on accuracy. We then tuned its

hyperparameters using PyCaret’s grid-search capabilities. The resulting model formed the baseline,

which achieved an accuracy of 40%.

5.1 BASELINE MODEL ON ITS OWN:

Before training the LightGBM classiﬁer, we employed the TF-IDF vectorizer from

scikit-learn to convert the tweet text into numerical features. This transformation allowed

the machine learning model to work with text data effectively.

For the baseline model, we performed grid search using the GridSearchCV function to tune the

hyperparameters of the LightGBM classiﬁer. The parameter grid included options for the number of

leaves, maximum depth, and number of estimators. After completing the grid search, we obtained

the best model based on the accuracy score, which achieved a classiﬁcation accuracy of 0.6905, or

69.05%. The classiﬁcation report (see Figure 2) provides a more detailed assessment of the model’s

performance. The report includes precision, recall, and F1-score metrics for each sentiment class,

as well as overall macro-averaged and weighted-averaged metrics.

Figure 2: Classiﬁcation Report For Baseline Model; Accuracy (Best Model): 0.6905; Best Pa-

rameters: {max depth: 20, n estimators: 200, num leaves: 30}.

Challenges faced during the construction of the baseline model are discussed in section 8: Project

Difﬁculty, of this document.

University of Toronto: APS360 Final Report – Group 21

6 QUANTITATIVE AND QUALITATIVE RESULTS

6.1 QUANTITATIVE RESULTS

The quantitative results of our model building were determined through a training and validation

process involving the larger dataset, with 828 data entries being split off to form a validation set and

the remaining 61692 entries comprising the training set. Despite a large difference in size between

the two datasets, an even split between all three sentiment classes, which would ensure that the

model will learn properly instead of making random guesses or “learning” to just predict the most

common class, as previous primary models were doing. In the training loop, each epoch, the model

would ﬁrst zero out all the gradients, performs a forward pass using each batch of inputs, calculates

loss using the CrossEntropyLoss function, calculates gradients, and then updates weights and biases.

After the training loop, validation would be conducted by performing a forward pass with our vali-

dation dataloader, including calculating accuracy and loss, but without updating weights and biases

after. The main quantitative results are the accuracy and loss for both training and validation, with

a step at each epoch to evaluate how they change over training. As shown below, training and val-

idation results both reached 96 percent accuracy while achieving a very small loss. Our highest

achieved validation accuracy was about 97.6

6.2 QUALITATIVE RESULTS

In terms of qualitative analysis, it would be useful to see how accuracy and loss changed with each

epoch. To get a good sense of this, graphs (see 3a, 3b) were created using the outputs of the training

loop to visualize the process. Confusion matrices were used to evaluate how the model would

perform on different ground truth labels. To understand where it was going wrong, it was important

to ﬁgure out how many guesses it was making for each possibility. The results of generating the

confusion matrix are shown below (see 3c). As the data shows, based on the training and validation

sets, the model performed very well on all 3 classes of data, with only a few misclassiﬁcations.

Also, the matrix shows that the model is not making the same mistake repeatedly, but just making

an error every so often. The graphs show that even from the ﬁrst epoch, the model was performing

well above a random guess or a repeated prediction in terms of accuracy. Regarding loss, training

loss can be seen to be extremely low due to the very large sample size. Validation loss appears to

increase in later epochs, which may suggest overﬁtting, but it still remains quite low.

(a) Accuracy Graph. (b) Loss Graph.

University of Toronto: APS360 Final Report – Group 21

7 EVALUATING MODEL ON NEW DATA: DELL TWEET DATA

7.1 DATA SOURCE AND PREPROCESSING CONSISTENCY

We meticulously ensured the integrity of our model evaluation by procuring a distinct dataset of

tweets exclusively targeting Dell. To guarantee an unbiased assessment, we conﬁrmed that this

dataset was entirely absent from our training and validation data. Moreover, we maintained the

same preprocessing pipeline, which included identical tokenization and padding techniques, as

employed during the training phase.

7.2 DEDICATED EVALUATION FUNCTION

In order to execute a comprehensive evaluation of our model on the Dell-speciﬁc dataset, we de-

vised a specialized function called predict and evaluate sentiment. This function was

meticulously designed to isolate the evaluation process, thus preventing any inadvertent inﬂuence

from hyperparameter tuning or prior training. By channeling our focus solely on the evaluation, we

adhered to best practices in avoiding data leakage and maintaining robustness.

7.3 MODEL EVALUATION AND PERFORMANCE

Our primary objective was to gauge the model’s generalization capability on the Dell-speciﬁc

dataset. This involved employing the specialized predict and evaluate sentiment func-

tion, which generated a confusion matrix. This matrix succinctly encapsulates the model’s pre-

dictive performance across the sentiment categories, showcasing areas of strength and areas that

required improvement.

7.4 RESULTS AND RELIABILITY

Upon conducting the evaluation, the obtained results are as follows:

Confusion Matrix:

6484 2825 1247

2456 2682 1910

1349 2817 3200

Accuracy - Negative: 61.42%

Accuracy - Neutral: 38.05%

Accuracy - Positive: 43.44%

Overall Accuracy: 49.52%

7.5 ENSURING MODEL INTEGRITY

The stringent measures we undertook to maintain the ﬁdelity of the evaluation process afﬁrm our

commitment to assessing the model’s performance on unseen data. By utilizing a separate dataset,

consistent preprocessing, and a dedicated evaluation function, we conﬁdently present a credible

representation of the model’s performance on new data, as outlined in the evaluation criteria.

8 DISCUSSION: DECIPHERING TWEET SENTIMENTS THROUGH DEEP

LEARNING INSIGHTS

8.1 MODEL PERFORMANCE AND INTERPRETATION

Our CNN architecture exhibited exceptional performance, achieving an accuracy of 97.6% on the

training and validation datasets. This accomplishment is particularly noteworthy given the intricate

nature of sentiment analysis. When extended to the Dell-speciﬁc dataset, our model demonstrated

a 50% accuracy, showcasing its adaptability to new data sources and its ability to generalize effec-

tively.

The model’s accuracy on the training, validation, and test datasets aligns consistently with our ex-

pectations. The confusion matrix on the validation data highlights the model’s proﬁciency in clas-

sifying sentiment categories. These results reﬂect the successful navigation of sentiment analysis

complexities.

Training Confusion Matrix:

260 2 4

4 277 4

1 5 271

8.2 CHALLENGES, INSIGHTS, AND FUTURE DIRECTIONS

Navigating data preprocessing challenges, including punctuation, contractions, and slang, required

a delicate balance between cleaning the tweets and retaining their essence. Striking this balance

was pivotal to the model’s performance. Future iterations could explore integrating pre-trained

word embeddings like GloVe to enhance the model’s vocabulary and address nuanced linguistic

variations.

University of Toronto: APS360 Final Report – Group 21

8.3 COMPARING MODEL ARCHITECTURES

Comparing, the baseline LightGBM model underscores the CNN architecture’s advantages. While

LightGBM provided prompt results, the CNN model excelled in handling sentiment nuances, result-

ing in improved accuracy. The CNN’s ability to capture intricate patterns and generalize effectively

makes it a powerful tool for sentiment analysis. However, it is important to note that the CNN ar-

chitecture’s computational complexity requires more resources compared to the LightGBM model.

8.4 EVALUATION ON DELL-SPECIFIC DATASET

The model demonstrated varying accuracy across sentiment classes on the Dell dataset, with 61.42%

accuracy for negative, 38.05% for neutral, and 43.44% for positive sentiments. Challenges in pre-

dicting neutral and positive sentiments shed light on the complexities inherent in these categories.

Accurate identiﬁcation of negative sentiments aligns seamlessly with the project’s objectives, as

it often signals issues requiring swift action. Predicting neutral and positive sentiments remains

intricate due to their diverse contexts and connotations.

8.5 LEARNING, REFLECTION, AND CONCLUSION

This project provided invaluable learning experiences, encompassing preprocessing, tokenization,

and deep model architectures. The iterative CNN model reﬁnement highlighted the value of perse-

verance and creative problem-solving. This sentiment analysis endeavor not only demonstrated the

prowess of deep learning but also underscored the signiﬁcance of robust preprocessing and model

design.

In conclusion, our model’s performance aptly matches the project’s complexity and expectations.

The attained accuracy rates validate the model’s capability in sentiment analysis, especially in dis-

cerning negative sentiments. Overcoming challenges throughout the project served as a testament to

the importance of meticulous preprocessing and thoughtful model architecture.

8.6 FUTURE DIRECTIONS AND BEYOND

Looking ahead, we envision enriching the model’s vocabulary by incorporating pre-trained word

embeddings, thus enhancing its contextual understanding. Tailoring models to industry-speciﬁc

linguistic nuances could further elevate sentiment prediction accuracy, making it a valuable tool

across diverse domains. We acknowledge the room for improvement for this model as well as deep-

learning practices that have come to light during the timeline of this project.

9 ETHICAL CONSIDERATIONS

Due to the ﬁeld in which this project is dealing, there are ethical considerations worth noting. Pri-

marily, the issue of data security and autonomy is an important factor when it comes to usage of this

network. The datasets that were used were publicly available on Kaggle, but if use of it were to be

expanded, it could be repurposed into a data scraping tool for social media platforms such as Twitter.

If a program parses through thousands of tweets, there may be problems with sensitive or private

information being given to companies, such as the ones that were being referenced in the tweet

datasets used. Even without explicitly sharing private information, in the process of data collecting,

it is possible to create a “dummy” tied to the account that holds the online information and behaviour

of a person for the purposes of targeted advertisement. While the model itself doesn’t have inher-

ent ethical issues, as it just classiﬁes inputs that can usually be classiﬁed by a person already, the

method by which additional data beyond the public datasets used could lead to questionable ethical

practices.

10 PROJECT DIFFICULTY AND PERFORMANCE EVALUATION

10.1 PROJECT DIFFICULTY ASSESSMENT

In this section, we assess the difﬁculty of the project and the performance of our model in relation

to the complexity of the problem. Our project involved tackling a large amount of data processing

and determining the correct build of the model. Despite the initial appearance of a simple language

classifying model, we identiﬁed several intricacies that contributed to the overall complexity of the

problem that had been mellowed down for our ﬁrst interaction with it, during the lecture.

1. Pre-processing Data: Dealing with edge cases in the raw data. There were multiple in-

stances that we realised required special consideration when we analyzed our raw data.

2. Baseline Model: Vectorization and its importance in ML and hyper-tuning.

University of Toronto: APS360 Final Report – Group 21

3. Deﬁning a Model: Striking the correct balance between complexity and simplicity for the

model to learn and perform

Our problem’s difﬁculty exceeded our expectations due to the interplay of these factors.

10.1.1 PRE-PROCESSING DATA

As mentioned earlier it proved to be a challenge for us to construct a data pipeline that accounted

for all data which also took into consideration multiple edge cases. As languages are ever changing

with the rise of slang, acronyms, contractions and all together evolution of the language, there were

an extensive amount of edge cases. Some encounters are elaborated below. We incorporated a

function that expands contraction and slang, but as we were using a dictionary that we collected and

constructed it only considered the most common cases to our knowledge.

Another challenge we faced was with punctuation. It is no doubt that punctuation can convey senti-

mental values [example: 1) Let’s eat, grandma! 2) Let’s eat grandma!]. Therefore we initially had

decided to keep the punctuation. However, later as we implemented the pipeline we realised that

ﬁrstly, it blew up tokenization and secondly, the padded sequence lengths grew too large. Therefore,

for the two reasons above we decided to remove all punctuation before we performed tokenization.

10.1.2 BASELINE MODEL

Initially, comprehending how to transform text data into numerical features was a complex task (vec-

torization of data for ML implementations). However, PyCaret’s built-in capabilities simpliﬁed this

process by automatically handling the vectorization of the tweet data. PyCaret’s preprocessing steps,

including the TF-IDF vectorizer, seamlessly converted the tweet text into numerical representations,

allowing the machine learning models to process the data effectively.

An additional challenge we faced was selecting the appropriate hyper-parameters for the Light-GBM

classiﬁer. To tackle this challenge, we employed grid search using PyCaret’s tuning capabilities. By

understanding the impact of these hyper-parameters on the model’s behavior, we were able to ﬁne-

tune the LightGBM classiﬁer and achieve optimal results for our baseline model.

10.1.3 DEFINING A MODEL

As mentioned in class that Sentiment Analysis is no easy feat, we came to realise this whilst de-

bugging our different models for the project. A difﬁculty we faced was evaluating why our initial

models maxed out performance at 60%. After quite some debugging we came to realise the trade-

off between a simple and a complex model. Our initial vanilla RNN, LSTM, and GRU maxed out

performance at approximately 30-35% because they were not able to learn as sentiment analysis is a

difﬁcult task. Our Bidirectional LSTM, GRU and quadruple stacked GRU with two fully connected

layers were either complex or too complex that they were only good a predicting one of the classes

deﬁnitely. It proved to be difﬁcult for us to strike a balance between our model being too simple

where to guessed labels for each epoch or was really good at determining one type of sentiment and

predicting that with a high accuracy.

10.2 MODEL PERFORMANCE EVALUATION

We strove to create a model that not only met the expectations of the project’s difﬁculty but also

demonstrated a level of performance that exceeded the anticipated challenges. Our model showcased

a performance that aligned with the intricacies of the problem, conﬁrming its capability to navigate

the complexities we outlined.

The model’s performance as seen by the matrix in section 6 during testing phases, substantiates

its effectiveness in addressing the problem’s intricacies. Although certain aspects of our model’s

performance like predicting the positive and neutral sentiments, were justiﬁably rooted in the unique

complexity of the problem. These limitations are temporary and as discussed in our future directions

section we strive to cure them to achieve better performances of our model.

10.3 LEARNING BEYOND EXPECTATIONS

Our team’s efforts extended beyond the teaching of the labs as we explored different implementa-

tions and libraries. By delving into the intricacies of the problem and implementing different models

to learn from, we enriched our understanding of Sentiment Analysis models through ﬁrst hand ex-

perience. Implementing various models heightened our understanding fop caveats of each as well as

exposed us to errors that we can now rectify and pin point to the exact part of the code that could be

malfunctioning. There was a lot of invaluable experience gained by pushing beyond our limits and

experimenting in the ﬁeld that we will carry forward in our journeys.

10.4 CONCLUSION

In conclusion, our project’s difﬁculty was undermined at ﬁrst but after implementing and getting our

hand dirty we began to realise the mammoth difﬁculty Sentiment Analysis poses.

University of Toronto: APS360 Final Report – Group 21

Our model’s performance, aligning with the identiﬁed challenges, although, it could be signiﬁcantly

improved. Through this project, we demonstrated our capacity to not only meet expectations but

also show perseverance in tough times by delving deep into the complexities and emerging with a

commendable solution.

Figure 4: Illustration

University of Toronto: APS360 Final Report – Group 21

REFERENCES

[1] Seiler, D.; “How COVID-19 has pushed companies over the technology tipping point-and

transformed business forever,” McKinsey & Company, https://www.mckinsey.com/

capabilities/strategy-and-corporate-finance/our-insights/how-

covid-19-has-pushed-companies-over-the-technology-tipping-

point-and-transformed-business-forever (accessed Aug. 10, 2023).

[2] A. Pandya and P. Lodha, “Social connectedness, excessive screen time during COVID-19 and

mental health: A review of current evidence,” Frontiers, https://www.frontiersin.

org/articles/10.3389/fhumd.2021.684137/full (accessed Aug. 10, 2023).

[3] Yadav, A., Vishwakarma, D.K. Sentiment analysis using deep learning architectures: a review.

Artif Intell Rev 53, 4335–4385 (2020). https://doi.org/10.1007/s10462-019-

09794-5

[4] G, Vinodhini, Chandrasekaran, Dr. (2012). Sentiment Analysis and Opinion Mining: A

Survey. Int J Adv Res Comput Sci Technol. 2. https://www.researchgate.net/

publication/265163299_Sentiment_Analysis_and_Opinion_Mining_A_

Survey

[5] L

aszl

o Nemes & Attila Kiss (2021) Social media sentiment analysis based on COVID-19,

Journal of Information and Telecommunication, 5:1, 1-15 https://doi.org/10.1080/

24751839.2020.1790793